Tensorflow选择GPU&CPU计算与训练==1

结果分析：Tensorflow搭载GPU与CPU计算与训练==2

1 默认模式

1.0 默认CPU

• 环境
• Ubuntu16.04
• python2.7
• Tensorflow1.12.0
• 源码Demo

v1 = tf.constant([125.0], shape=[1], name='v1')
v2 = tf.constant([125.0], shape=[1], name='v2')
result = v1 + v2
# 输出使用设备的信息True
config = tf.ConfigProto(log_device_placement=True)
with tf.Session(config=config) as sess:
	print("GPU calculator result is : {}".format(sess.run(result)))
	print("CPU calculator time consuming is: {}".format(timeConsuming))

1.2 默认GPU

• 环境
• Ubuntu16.04
• python2.7
• Tensorflow-gpu 1.4.0
• CUDA 8
• cuDNN 6
• 源码Demo

v1 = tf.constant([125.0], shape=[1], name='v1')
v2 = tf.constant([125.0], shape=[1], name='v2')
result = v1 + v2
# 输出使用设备的信息True
config = tf.ConfigProto(log_device_placement=True)
with tf.Session(config=config) as sess:
	print("GPU calculator result is : {}".format(sess.run(result)))
	print("GPU calculator time consuming is: {}".format(timeConsuming))

2 指定模式

2.1 全局指定

2.1.0 全局CPU

• 环境
• Ubuntu16.04
• python2.7
• Tensorflow1.12.0
• 源码Demo

import os 
# 定义放在导入tensorflow前
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
import tensorflow as tf
# CPU default
v1 = tf.constant([125.0], shape=[1], name='v1')
v2 = tf.constant([125.0], shape=[1], name='v2')
startTime = time.time()
result = v1 + v2
endTime = time.time()
timeConsuming = endTime - startTime
with tf.Session(config=tf.ConfigProto(log_device_placement=True)) as sess:
	print("CPU calculator result is : {}".format(sess.run(result)))
	print("CPU calculator time consuming is: {}".format(timeConsuming))

2.1.2 全局GPU

• 环境
• Ubuntu16.04
• python2.7
• Tensorflow-gpu 1.4.0
• Tensorflow 1.12.0
• CUDA 8
• cuDNN 6
• 源码Demo

import os
# 全局指定GPU编号
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
import tensorflow as tf
# GPU default
v1 = tf.constant([125.0], shape=[1], name='v1')
v2 = tf.constant([125.0], shape=[1], name='v2')
startTime = time.time()
result = v1 + v2
endTime = time.time()
timeConsuming = endTime - startTime
with tf.Session(config=tf.ConfigProto(log_device_placement=True)) as sess:
	print("GPU calculator result is : {}".format(sess.run(result)))
	print("GPU calculator time consuming is: {}".format(timeConsuming))

2.2 局部指定

2.2.1 局部CPU

• 环境
• Ubuntu16.04
• python2.7
• Tensorflow-gpu 1.4.0
• Tensorflow 1.12.0
• CUDA 8
• cuDNN 6
• 源码Demo

with tf.device('/cpu:0'):
	v1 = tf.constant([125.0], shape=[1], name='v1')
	v2 = tf.constant([125.0], shape=[1], name='v2')
	result = v1 + v2
	# 输出使用设备的信息True
	config = tf.ConfigProto(log_device_placement=True)
	with tf.Session(config=config) as sess:
		print("CPU calculator result is : {}".format(sess.run(result)))
       print("CPU calculator time consuming is: {}".format(timeConsuming))

2.2.2 局部GPU

• 环境
• Ubuntu16.04
• python2.7
• Tensorflow-gpu 1.4.0
• Tensorflow 1.12.0
• CUDA 8
• cuDNN 6
• 源码Demo

with tf.device('/gpu:0'):
	v1 = tf.constant([125.0], shape=[1], name='v1')
	v2 = tf.constant([125.0], shape=[1], name='v2')
	result = v1 + v2
	# 输出使用设备的信息True
	config = tf.ConfigProto(log_device_placement=True)
	with tf.Session(config=config) as sess:
		print("GPU calculator result is : {}".format(sess.run(result)))
       print("GPU calculator time consuming is: {}".format(timeConsuming))

3 总结

• 依据不同需求,选择不同环境;
• GPU与CPU可全局配置也可局部配置,多GPU可选择局部配置;
• CPU与GPU计算速度不同(废话),单纯CPU环境与GPU环境下的CPU处理效率是不同的,参见另一篇文章.
• Tensorflow选择GPU&CPU计算与训练==2

---

[参考文献]
[1]https://blog.csdn.net/qq_35559420/article/details/81460912
[2]https://blog.csdn.net/dcrmg/article/details/79747882

---